The GaaS Pricing Report: What 50 Agentic AI Vendors Actually Charge

Table of Contents

• Why a Benchmark, and Why Now
• How We Built the Sample
• The Headline Distribution
• Per-Outcome Pricing: The Numbers Behind the Slogan
• Per-Task and Metered: The Quiet Default
• Per-Seat Survivors and the Hybrids
• What Vendors Hide on the Pricing Page
• Vertical Pricing Premiums
• Insights Most People Overlook
• References

Why a Benchmark, and Why Now

Every founder pitching an agent product right now will tell you their pricing is "aligned to value." Almost none of them can tell you what the vendor down the hall charges for a comparable unit of work. The category is young enough that there is no Rule of 40, no SaaS Capital index, no shared vocabulary for what a "good" gross margin looks like when a chunk of your COGS is somebody else's inference bill.

That gap is the reason this report exists. Over six weeks we collected published and quoted pricing from 50 vendors selling agents as a service, not copilots, not chat UIs, but systems that take an instruction and run a multi-step workflow to completion with limited human babysitting. The point was not to crown a winner. It was to answer a question buyers keep asking and analysts keep dodging: when the marketing says "outcome-based," what does the invoice actually say?

The short version is that the market is messier and more honest-by-accident than the keynotes suggest. The slogans have converged on outcomes. The billing has not.

How We Built the Sample

We are upfront about method because pricing benchmarks live or die on it. The 50 vendors span six segments: customer support, sales and SDR automation, software engineering, recruiting and HR, finance and back-office, and horizontal "build-your-own-agent" platforms. We required each to be selling to businesses, charging real money, and operating agents with at least three chained steps of autonomy. Pure LLM API resellers were excluded; so were single-shot summarizers dressed up as agents.

Pricing came from three sources, in order of trust: public pricing pages, signed quotes shared by buyers under anonymity, and sales-call notes where a rep stated a number on the record. Where a vendor only does "contact sales," we logged that as a data point in itself, opacity is a pricing strategy. All figures were normalized to a common unit per segment (a resolved ticket, a sourced candidate, a merged pull request) so the comparisons mean something. Inference-cost assumptions were held constant using late-2025 frontier-model rates, which the major providers publish openly, including Anthropic's model pricing documentation. This is a snapshot, not a moving average; in a market repricing this fast, anything older than a quarter is fiction.

The Headline Distribution

Here is the spread that surprised even us. Of the 50:

• 18 vendors lead with usage or per-task metering as the primary model.
• 16 vendors still anchor on per-seat or flat-tier subscriptions, sometimes with a usage rider bolted on.
• 11 vendors advertise per-outcome or per-resolution pricing as the headline.
• 5 vendors run a genuine hybrid where a base platform fee and a usage/outcome component are both first-class, not one buried under the other.

The takeaway that matters: outcome pricing dominates the messaging but not the billing. More than three-quarters of these companies still collect most of their revenue through a seat or a meter. The "we only charge when it works" banner is frequently sitting on top of a usage meter doing the actual accounting, a tension we dig into in the disclosure section. McKinsey's work on the economic potential of generative AI frames the value at stake, but value framing and billing mechanics are not the same animal, and vendors blur them constantly.

Per-Outcome Pricing: The Numbers Behind the Slogan

The most-cited reference point in the category is support. Intercom's Fin set an anchor at $0.99 per resolution, and that single number has done more to shape buyer expectations than any analyst report. Across our support sample, headline per-resolution prices clustered between $0.65 and $1.50, with a median right around $0.99, vendors are quite literally pricing to the Fin anchor, whether they admit it or not.

Outside support, outcome pricing gets fuzzier because the outcome gets harder to define. In recruiting, "per qualified candidate sourced" ran $8 to $40 depending on seniority of the role, but three of the five vendors quoting that model reserved the right to bill for candidates the customer later rejected, a definitional landmine we keep seeing across the category. In sales, "per meeting booked" showed up at $15 to $75, and the disputes there are entirely about what counts as a meeting that happened.

The pattern is consistent: the cleaner the outcome is to observe, the more credible the outcome price. A resolved support ticket has a timestamp and a CSAT score. A "qualified lead" has a salesperson's opinion. That observability gradient, more than any pricing philosophy, predicts whether outcome billing survives contact with a real contract.

Per-Task and Metered: The Quiet Default

If you ignore the marketing and just look at where revenue accrues, metered per-task pricing is the real center of gravity. Eighteen vendors lead with it, and several of the "outcome" vendors fall back to it under the hood. The unit varies, per run, per workflow execution, per thousand "agent actions," per credit drawn from a prepaid pool, but the economics rhyme.

Coding agents were the cleanest example. Per-task pricing for an autonomous code change landed between $2 and $20 per accepted task, with the spread driven almost entirely by how many model calls a task chews through. This is where margin discipline lives or dies. The vendors with healthy unit economics were the ones routing cheap models for easy steps and reserving frontier models for the hard ones, a technique a16z has flagged in its writing on the cost structure of AI applications. The vendors bleeding margin were the ones passing every step to the most expensive model and eating the difference to keep the price simple.

The honest appeal of metered pricing is that it scales with the work and protects the vendor's margin. The honest problem is buyer anxiety: a meter that ticks while nobody is watching is a budgeting nightmare, which is exactly why prepaid credit pools and hard usage caps have spread so fast as a pressure-release valve.

Per-Seat Survivors and the Hybrids

The most contrarian finding in the dataset is that per-seat pricing is not dead. Sixteen vendors still anchor on it, and not because they are behind the times. Per-seat survives wherever the buyer is an enterprise procurement team that wants a predictable line item, and wherever the agent augments a named human rather than replacing a unit of output.

But "per seat" increasingly means something new. Several vendors price a seat that includes a generous bundle of agent actions, then meter overage above it, a base-plus-usage hybrid wearing a seat costume. Five vendors run an explicit hybrid with both a platform fee and a usage component as headline numbers, and those were, notably, the ones with the most stable revenue and the fewest pricing-page complaints in buyer interviews. The base fee buys predictability; the usage component captures expansion. When expansion is automatic, usage just grows as the agent does more work, that hybrid quietly becomes one of the strongest land-and-expand motions in software.

What Vendors Hide on the Pricing Page

This is the section buyers should read twice. The gap between the headline price and the all-in price is where the category's worst behavior lives.

The most common omission is the floor. Eleven vendors advertising per-outcome or per-usage pricing carried an undisclosed minimum monthly commitment, typically $2,000 to $10,000, that only surfaced on a sales call. "Pay only when it works" is technically true and practically misleading when there is a five-figure floor underneath it.

The second is outcome definition. Almost no outcome-priced vendor publishes who adjudicates the outcome or how disputes get resolved. When the vendor both performs the work and grades it, the incentive to grade generously is obvious, and we saw real customer complaints about resolutions that "resolved" nothing.

The third is inference pass-through volatility. A handful of metered vendors reserve the right to adjust per-unit prices as their model costs change, usually framed as a customer benefit, occasionally used as cover for a quiet increase. Gartner's analysis of how generative AI pricing is unsettling software budgets underscores why FinOps teams have started treating agent spend like cloud spend: assume it drifts, and cap it contractually.

The fourth, and the one that should be a dealbreaker, is the absence of a failure SLA. If an agent fails the task, what happens? Most pages are silent. The vendors worth trusting spell out refunds or credits for failed and partially completed work; the rest leave it to the dispute you will eventually have.

Vertical Pricing Premiums

Vertical agents charge more, and mostly they should. A support agent and a finance-reconciliation agent might run identical token counts under the hood, but the finance agent carries audit overhead, compliance liability, and a far higher cost of being wrong. Across our sample, agents in regulated domains, finance, healthcare, legal, priced 40% to 120% above the horizontal equivalent for a comparable unit of work.

The premium is not pure rent. Regulated-industry agents genuinely cost more to run: human-in-the-loop review steps, audit-log retention, narrower model choices for data-residency reasons. But the premium is also where vendors with proprietary data and workflow lock-in extract the most, because outcome pricing structurally favors the incumbent who already owns the customer's data. The buyer's job is to separate the part of the premium that buys real risk reduction from the part that buys the vendor's moat.

Insights Most People Overlook

The "outcome" slogan is a marketing layer, not a billing layer. More than three-quarters of vendors waving the outcome banner collect most revenue through a seat or meter. Read the invoice, not the homepage. The convergence everyone narrates is happening in copywriting, not in accounting.

Undisclosed floors quietly invert the risk story. "Pay only when it works" plus a $5,000 monthly minimum is a subscription with extra steps. The floor, not the headline rate, determines whether a small buyer can afford the product, and it is almost never on the public page.

Whoever defines the outcome owns the margin. The single most important clause in an outcome contract is not the price; it is the adjudication mechanism. A vendor that performs and grades its own work has a structural incentive to grade loosely. Demand an auditable definition before you sign.

Model routing, not pricing model, separates the survivors. Two vendors can charge the identical per-task price and have wildly different gross margins purely based on which model handles which step. Pricing strategy gets the keynote; routing strategy pays the rent. When inference costs drop, the routers expand margin while the rest face the grandfather problem of repricing.

Opacity correlates with weakness, not strength. The vendors hiding behind "contact sales" were disproportionately the ones with shaky unit economics or undefined outcomes. The strongest performers, the explicit hybrids, were also the most transparent, because predictable economics are easy to publish. Treat a hidden price as a signal, not just an inconvenience.

References

• Anthropic, Models Overview and Pricing
• McKinsey, The Economic Potential of Generative AI: The Next Productivity Frontier
• a16z, The Economics of Generative AI
• Gartner, Take This View to Rein In Generative AI Costs